CN112839746B - Folding produces prediction system - Google Patents

Abstract

The fold generation prediction system collects and stores the first data indicating the presence or absence of fold generation and the generation location in the target rolling pass, and the first data including the rolling material that is earlier in the rolling order than the target rolling pass. The information related to the previous rolling pass and the second data of the properties related to the rolling material at the time of the previous rolling pass are used as data for building an adaptive model. The folding generation prediction system constructs an adaptive model using the saved adaptive model building data, and saves the adaptive completion model that has been constructed. The folding generation prediction system collects the information related to the preceding rolling pass when the rolling material containing the prediction object is rolled earlier than the object rolling pass in the rolling sequence, and the information related to the preceding rolling pass. Data for prediction of properties related to the rolled material. The fold occurrence prediction system inputs the data for prediction into the adaptive completion model, and predicts whether or not folds are generated in the target rolling pass and all the occurrence parts before the rolling material to be predicted reaches the target rolling pass. or part of it.

Description

Folding produces prediction system

technical field

The present invention relates to a system for predicting in advance that a fold (Japanese: twist) will occur in a rolled material during hot rolling in which a plate-shaped metal material is heated to a high temperature and rolled through a plurality of rolling passes.

Background technique

Rolling mills thin blocks of non-ferrous materials such as steel, aluminum, and copper by rolling them so that they can be easily processed into automotive and electrical products. There are various types of rolling mills, such as a hot sheet rolling mill for rolling sheets, a heavy plate rolling mill, a cold rolling mill, and a rolling mill for rolling rods and wires. Among them, the hot sheet rolling machine that rolls the rolled material in batches at a high speed in batches is prone to folds.

FIG. 12 is a diagram showing an example of the configuration of a rolling mill in a conventional hot sheet rolling process. The rolling mill 20 shown in FIG. 12 is constituted by various devices such as a heating furnace 21 , a roughing mill 22 , a bar heater 24 , a finishing mill 25 , a delivery table 26 , and a coiler 27 . The rolled material 100 heated by the heating furnace 21 is rolled by the rough rolling mill 22 . The rolled material 100 rolled by the roughing mill 22 is conveyed to the finishing mill 25 via the bar heater 24 . After the rolling material 100 rolled by the finishing mill 25 is cooled on the delivery table 26 , it is wound into a coil shape by the coiler 27 . A coil-shaped sheet obtained by thinly rolling the rolled material 100 is the final product.

The roughing mill 22 shown in FIG. 12 includes: a rolling stand R1, each having an upper and lower work roll; and a rolling stand R2, including a total of 4 upper and lower work rolls and upper and lower backup rolls having a diameter larger than that of the upper and lower work rolls root roll. The finishing mill 25 shown in FIG. 12 has seven rolling stands F1 to F7 arranged in series. In the example shown in FIG. 12 , the rolling stands F1 to F7 of the finishing mill 25 are composed of four upper and lower rolls, but there are also six upper and lower rolls including intermediate rolls that enter between work rolls and backup rolls. composition situation. In addition, the structure of the apparatus is similar in many cases, although the large-capacity motors for driving the upper and lower rolling rolls, the shafts connecting the rolls and the motors, etc. differ in minute specifications.

Side guides (not shown) are provided on the entrance sides of the respective rolling stands of the roughing mill 22 and the finishing mill 25 . In the rough rolling mill 22, the rolling material is often stopped before the rolling of the material, the opening degree of the side guides is narrowed to sandwich the rolling material, and the rolling is performed after centering. In the finishing mill 25, since the rolling material often enters the rolling stand at a high speed, the opening degree of the side guides is often set in advance by the width obtained by adding the allowance to the width of the rolling material.

Folding of the rolling material is a phenomenon in which the rolling material meanders directly under the rolling stand, that is, the leading or trailing end of the rolling material due to movement in the roll width direction or bending in the width direction. The folding includes a leading end fold generated at the leading end of the rolled material and a trailing end fold generated at the trailing end of the rolled material. Front end folding is caused by the meandering of the rolling material and the bending of the front end of the rolling material, so that the front end of the rolling material hits the entrance side guide plate before entering the rolling stand, and enters the rolling stand while bending the front end. produced. The tail end is folded because the tail end of the rolling material meanders before leaving the rolling stand, and the tail end collides with the entry side guide plate, or the tail end is bent into two sheets and rolled at the same time, so that the load is concentrated at the same time. The bent part, the part is torn, etc.

If folding occurs, the roller surface will be damaged. In order to prevent this damage from being transferred to the surface of the next rolling material, the operation may be temporarily stopped, and the rolls may be pulled out for inspection. In addition, the fractured end of the fractured material may remain in the rolling mill. Since this fractured edge may hinder the passability of the material to be rolled next, inspection is also required in such a case. These operations reduce productivity and even reduce roll unit consumption (Japanese: ロール original unit).

Also, in the rough rolling mill 22, so-called reversing rolling is performed by repeating forward and reverse rolling, and there are 5 and 6 passes, and in the finishing rolling mill 25, there are 6 or 7 rolling stands F1 to F7. Rolled through the air. One pass of the rolled material under the rolling stand is referred to as one pass. In the rough rolling mill 22, a plurality of passes of rolling are performed by one rolling stand, and in the finishing rolling mill 25, only one pass of rolling is performed by one rolling stand. Hereinafter, the trailing end folding in the finishing mill 25 in which the frequency of occurrence of folding is particularly high will be described. Here, rolling in one pass is the same as rolling in one stand.

In the conventional hot sheet rolling process, the following countermeasures are generally taken against folding.

Countermeasure A: For rolled materials that are prone to folds, operators should handle them in advance.

Countermeasure B: Respond to the situation where folding occurs, and the operator immediately responds.

Countermeasure C: Apply automatic meandering control that suppresses meandering at the tail end.

Among the rolled materials that are prone to folds, there are rolled materials with thin product thickness, rolled materials with small plate crowns, specific steel grades, and the like. In particular, folds tend to occur in the latter stage of the finishing mill where the sheet thickness is thin and the rolling speed is increased. According to the countermeasure A, with respect to such a rolling material and a situation, the operator handles it while observing the meandering state of the rolling material on the upstream side. However, the operator has to deal with various situations such as the pressing force and the speed, and since the level of proficiency of each operator is different, it is not always possible to deal with it accurately. In the countermeasure B, when the meandering starts, the operator corrects it. However, since the meandering is a phenomenon that occurs rapidly, the operator may not be able to deal with it accurately.

Countermeasure C is effective in suppressing the meandering, but it is controlled after the meandering starts, rather than predicting the occurrence of meandering in advance. In addition, as disclosed in, for example, Patent Document 1 and Patent Document 2, meandering control has been performed for a long time. Regarding the prior art disclosed in Patent Documents 1 and 2, although the specific methods thereof are different, the meandering amount of the rolling material is calculated and controlled to suppress the meandering amount and prevent folding. However, the object of snake control is an unstable system, which is difficult to control and there is no effective control method.

prior art literature

Patent Literature

Patent Document 1: Japanese Patent Laid-Open No. 2018-43255

Patent Document 2: Japanese Patent Application Laid-Open No. 4-118108

SUMMARY OF THE INVENTION

The problem to be solved by the invention

Whether and where folds occur in the rolled material rolled in front of the operator's eyes is largely dependent on the operator's experience and intuition, and it is difficult to accurately predict them. Although attempts have been made to construct the meandering of the board as a cause of folding with a physical model, it has been difficult to construct a model with sufficient accuracy in reality. In addition, it is difficult to say that sufficient performance is obtained using the meandering control of the model.

The present invention has been made in view of such a problem, and an object of the present invention is to provide a fold occurrence prediction system capable of predicting the presence or absence of fold occurrence and its occurrence location.

means to solve the problem

The fold generation prediction system of the present invention predicts the generation of folds due to meandering of the rolled material or in the width direction in hot rolling in which a sheet-like metal material is heated to a high temperature and rolled through a plurality of rolling passes A phenomenon that occurs at the leading or trailing end of the rolled material due to upward bending, wherein the system is provided with one or more computers. One or more computers are programmed to perform the following processes: the process of collecting and saving adaptive model building data used in the construction of the adaptive model used to predict the generation of folds; building using the adaptive model The process of using the data to construct an adaptive model; the process of saving the adaptive model that has been constructed, that is, the adaptive completion model; the process of collecting the prediction data used in the prediction generated by the folding; and by inputting the prediction data to An adaptive completion model to predict the generation of folds.

In detail, in the process of collecting and saving data for adaptive model construction, one or more computers provided in the folding generation prediction system collect multiple sets of first data and second data as data for adaptive model construction. The first data indicates the presence or absence of the occurrence of folds in the target rolling pass that is the target of the fold occurrence prediction, and the position of occurrence of the folds, and the second data includes the previous rolling pass of the rolled material associated with the first data. Information related to the preceding rolling pass that precedes the target rolling pass in the rolling sequence and properties related to the rolled material at the time of the secondary rolling. In the process of collecting the data for prediction, when the rolling material including the prediction target is rolled in the preceding rolling pass that is earlier than the target rolling pass in the rolling order, the difference between the preceding rolling and the preceding rolling is collected. Information on the pass and data on properties related to the rolled material are used as prediction data. In the process of predicting the occurrence of folds, before the rolling material to be predicted reaches the target rolling pass, the presence or absence of occurrence of folds in the target rolling pass and all or part of the occurrence site are predicted.

One or more computers included in the folding generation prediction system may be programmed to perform processing of displaying the prediction result of folding generation on the display device.

One or more computers included in the fold occurrence prediction system may be programmed to perform processing for operating the entry-side guide of the target rolling pass when folds are predicted to be generated in the target rolling pass. In the process of operating the entrance-side guide, it is possible to determine which end of the front end and the trailing end of the rolling material to be predicted is folded, and to open the entrance-side guide in accordance with the passage of the folded end. In addition, it is also possible to determine which side of the operation room side and the motor side of the target rolling pass is folded, and to open the inlet side guide on the side where the fold occurs. When it cannot be determined which side of the operation room side and the motor side of the target rolling pass is folded, the inlet side guides on both sides of the operation room side and the motor side may be opened.

In the process of constructing the adaptive model, the adaptive model can also be constructed by machine learning or statistical methods included in the category of artificial intelligence, and the adaptive model is updated every time a certain amount of data for constructing an adaptive model is newly obtained.

In the process of collecting and saving the data for building an adaptive model, the presence or absence and occurrence of folds in the target rolling pass may be determined by analyzing the image data of the rolled material that has passed the target rolling pass. part. Alternatively, based on the load applied to the entry-side guide plate of the target rolling pass, the presence or absence of the occurrence of folds in the target rolling pass and the occurrence position may be determined. In addition, the presence or absence of the occurrence of folds in the target rolling pass input by the operator via the HMI and the occurrence location can also be received.

Invention effect

According to the folding generation prediction system of the present invention, data including information on the presence or absence of folding generation prediction of rolling materials are collected, and an adaptive model is constructed by using machine learning and statistical methods, and the rolling pass of the prediction object is upstream. The data is input to the adaptive completion model, so that the presence or absence and location of folds in the material to be rolled can be predicted in advance. Thereby, it is possible to have a time margin for preparations such as preventing or reducing the meandering of the plate, which is a cause of folding, before the rolling material passes through the target rolling pass, so that the occurrence of folding can be reduced and the resulting stability can be achieved. The operation can further lead to an improvement in the consumption per roll of the roll. In addition, since the prediction is made based on the actual machine data, there is an advantage of following changes in the situation of the actual machine.

Description of drawings

FIG. 1 is a block diagram showing the configuration of a fold occurrence prediction system according to the first embodiment of the present invention.

FIG. 2 is a diagram showing an example of an HMI used by an operator to input the presence or absence of folding generation and the generation location.

3 is a diagram conceptually showing processing performed by an adaptive model building unit of the fold generation prediction system according to the first embodiment of the present invention.

FIG. 4 is a diagram conceptually showing the processing performed by the prediction unit of the fold occurrence prediction system according to the first embodiment of the present invention.

FIG. 5 is a flowchart showing in detail the processing flow of the folding occurrence prediction system according to the present embodiment.

FIG. 6 is a diagram illustrating a configuration example of an SOM and an outline of a prediction method using the configuration example of the SOM.

FIG. 7 is a diagram for specifically explaining the outline of the operation of the ACC.

FIG. 8 is a diagram specifically explaining the outline of the operation of the ACC.

FIG. 9 is a diagram specifically explaining the outline of the operation of the ACC.

FIG. 10 is a flowchart showing a modification of the processing flow of the folding occurrence prediction system according to the present embodiment.

FIG. 11 is a diagram showing an example of control of the entrance side guide using the fold prediction result of the fold occurrence prediction system according to the second embodiment of the present invention.

FIG. 12 is a diagram showing an example of the configuration of a rolling mill in a conventional hot sheet rolling process.

Detailed ways

Embodiments of the present invention will be described with reference to the drawings. However, the embodiments shown below illustrate apparatuses and methods for embodying the technical idea of the present invention, and are not intended to limit the configuration, arrangement, processing order, etc. of the components to the following unless otherwise specified. content. The present invention is not limited to the embodiments shown below, and can be implemented with various modifications within a range that does not deviate from the gist of the present invention.

<First Embodiment>

1. The composition of the folding generation prediction system

FIG. 1 is a block diagram showing the configuration of a fold occurrence prediction system according to the first embodiment of the present invention. The folding generation prediction system of the present embodiment is applied to a hot sheet rolling process having the configuration shown in FIG. 12 . The folding occurrence prediction system includes a prediction device 10 that collects and stores predicted data, and performs calculations using the data. The prediction apparatus 10 may be constituted by a single computer, or may be constituted by a plurality of computers connected to a network.

The prediction apparatus 10 includes an adaptive model construction data acquisition and storage unit 1 , an adaptive model construction unit 2 , a prediction data acquisition unit 3 , a prediction unit 4 , and a result display unit 5 . Among these elements, the adaptive model construction data acquisition and storage unit 1 , the adaptive model construction unit 2 , the prediction data acquisition unit 3 , and the prediction unit 4 are executed by the processor by executing a program read from the memory of the computer. Implemented in software. Various programs and various data used for folding generation prediction are stored in the memory. In addition, the memory mentioned here includes both the main storage device and the auxiliary storage device. The result display unit 5 is a display device combined with a computer.

The adaptive model construction data acquisition and storage unit 1 acquires and stores adaptive model construction data for constructing an adaptive model to be described later. The first data and the second data are included in the adaptive model building data. The first data is data indicating the presence or absence of the occurrence of folds in the rolling stand (rolling pass) targeted for the prediction of the occurrence of folds, and the location of occurrence of folds. The folded portion is classified into the front end and the rear end in the flow direction (length direction) of the rolling material, and is classified into the work side (Work Side: hereinafter, referred to as WS) and the motor side ( DriveSide: hereinafter, referred to as DS). In the first data, the identification number (ID) of the rolling material rolled by the target rolling stand is associated with product information such as plate thickness and plate width, and is stored together with the first data.

The second data is process data including information about the preceding rolling stand when the rolling material is rolled by the preceding rolling stand, and properties about the rolling material. The stand precedes the object rolling stand in the rolling sequence. The information on the preceding rolling stand includes, for example, information on data items such as roll gap, roll shift amount, rolling load, thickness gauge, and the like, which can be acquired by sensors. In a finishing mill having a total of n rolling stands, the target rolling pass for predicting the occurrence of folds, that is, the target rolling stand can be, for example, from the m-th stand to the final n-th stand (m≤ n) rack. However, in the prediction stage, since it is not clear which rolling stand is designated as the target rolling stand, the data of all the stands of the finishing mill are collected and stored. In addition, the properties related to the rolling material refer to the properties of the rolling material rolled by the previous rolling stand, and refer to the identification number, steel type, target plate thickness, target plate width, and target plate of the rolling material. Convexity, target flatness, target temperature, etc. The first data and the second data are associated via the identification number of the rolling material.

The adaptive model construction data acquisition and storage unit 1 is combined with a not-shown HMI (Human-Machine Interface). The presence or absence of folding generation and the generation location are included in the first data of the adaptive model building data. These can be entered by the operator via the HMI. The site of fold generation is generally the WS or DS in the leading or trailing end. However, it is not always easy for an operator to visually determine whether it is WS or DS. Therefore, in order to absorb the uncertainty of the visual observation, in addition to the WS and DS, a generation location near the center of the width (Center of width: hereinafter, referred to as CW) may be provided. In addition, in order to classify and determine the generation site in detail, it is necessary to input a very large amount of data to the adaptive model, and in this case, the construction of the adaptive model takes a lot of time. Therefore, for example, only the leading end and the trailing end are classified and the like, allowing the generation site to be roughly divided.

FIG. 2 is a diagram showing an example of an HMI used by an operator to input the presence or absence of folding generation and the generation location. WS and DS shown in the figure are buttons, and are provided for each of the seven rolling stands F1 to F7 and each of the front end side (Head) and the tail end side (Tail). For example, when the WS at the rear end of the third rolling stand F3 is visually observed to be folded, the operator presses the WS button in the row of Tail in the column F3. In addition, in the HMI shown in FIG. 2, the CW button for absorbing the uncertainty of visual observation may be added. Alternatively, instead of adding the CW button, when there is a fold and it is uncertain which of the WS and DS buttons is, pressing both the WS button and the DS button may perform the same judgment as pressing the CW button.

As a method of acquiring the first data for constructing an adaptive model, in addition to the operator inputting the presence or absence of folding and the location of occurrence, it is also possible to use a method of determining from image data. Generally, there are several TV cameras in the rolling mill. By photographing the rolling material passing through the rolling stands with a television camera installed between the rolling stands of the finishing mill, and analyzing the obtained image data, it is possible to easily determine whether the rolling material is at the leading end or the trailing end. folded. When a fold occurs, the rolled material is torn, and the high-temperature portion inside the material appears to be streaked. Since the surface temperature of the rolled material is lower than that of the inside and it is black, the high temperature part inside the fractured part is clearly seen as orange. In addition, based on the image data, it is also easy to determine whether the portion that collides with the side guide is WS or DS.

As another method of acquiring the first data for building an adaptive model, a method of determining by the load applied to the side guide plate of the rolling stand can also be used. The side guide can be controlled by position or force, and the force applied to the side guide can be detected by a sensor. Folds are often caused by meandering of the rolling material, and when the force applied to the side guides is greater than or equal to a certain threshold value, it can be determined that folds have occurred. In addition, since the force applied to the side guides can be independently detected on the left and right, it can also be determined in which of the WS and the DS the folding occurs.

Returning to Fig. 1 again, the description of the configuration of the folding generation prediction system is continued. When a certain amount of data is collected by the adaptive model construction data collection and storage unit 1 , the adaptive model construction data collection and storage unit 1 inputs the collected and stored data to the adaptive model construction unit 2 . The adaptive model constructing unit 2 constructs an adaptive model 2a using the input data. An adaptive model refers to a model in which the relationship between internal components changes and the output changes when data is input.

As an example of an adaptive model suitable for prediction by folding, a neural network (Neural Network: hereinafter, referred to as NN) or Self Organizing Map (hereinafter, referred to as SOM) included in the category of machine learning can be cited. etc., and Adaptive Control Charts (Adaptive Control Charts: hereinafter, referred to as ACC) using statistical methods, etc. There are two types of learning methods classified as methods of machine learning, teacher learning and teacherless learning. In general, NN performs teacher-teachable learning, and SOM performs teacher-less learning. They can also be applied to two-valued problems classified as presence or absence of folds. Methods like NN, SOM, ACC are well known methods. The outline of these methods will be described later.

FIG. 3 is a diagram conceptually showing processing performed by the adaptive model building unit 2 . The adaptive model constructing unit 2 inputs the second data 301 for constructing an adaptive model to the adaptive model 302 (the adaptive model 2a shown in FIG. 1 ). Thereby, the degree of relationship between the internal components changes, and the adaptive model 302 is constructed. In addition, the adaptive model construction unit 2 inputs the first data 303 for adaptive model construction into the adaptive model 302 as teacher data or verification data. When the first data 303 is used as the teacher data, the difference between the output of the adaptive model 302 obtained from the second data and the first data 303 is returned to the adaptive model 302 as backpropagation.

Returning to Fig. 1 again, the description of the configuration of the folding generation prediction system is continued. After the adaptive model 2 a is constructed, the adaptive model constructing unit 2 stores the constructed adaptive model 2 a separately as the adaptive completed model 4 a and the adaptive model 2 a. The stored adaptive completion model 4a is used in the prediction section 4 for the prediction of the folded generation. The reason why the adaptive completion model 4a is stored separately from the adaptive model 2a is that the internal state of the adaptive completion model cannot be changed during the generation of the prediction fold.

The prediction data collected by the prediction data collection unit 3 is used for prediction using the folding of the adaptive completion model 4a. The prediction data acquisition unit 3 acquires the same type of data as the second data for adaptive model construction as prediction data. That is, the prediction data acquisition unit 3 sequentially changes the rolling stands of the prediction object from the m-th stand to the final n-th stand, and collects the rolling material including the prediction object to be rolled by the preceding rolling stand. information on the preceding rolling stand, which is earlier than the rolling stand to be predicted, and data on attributes related to the rolling material to be predicted. First. The prediction data acquisition unit 3 inputs the acquired prediction data to the adaptive completion model 4a.

FIG. 4 is a diagram conceptually showing the processing performed by the prediction unit 4 . The prediction unit 4 inputs the prediction data 311 to the adaptation completion model 312 (the adaptation completion model 4a shown in FIG. 1 ), and obtains the prediction result of the presence or absence of folding as the output 313 of the adaptation completion model 4a. In addition, when there is a fold, the prediction result of the generation part is also obtained as the output 313 of the adaptive completion model 4a. The occurrence sites of folds obtained as a result of the prediction are not necessarily all occurrence sites, and may be a part of the occurrence sites. In addition, the prediction using the adaptive completion model 312 is performed so that the prediction result is obtained before the rolling material to be predicted reaches the target rolling stand.

Returning to Fig. 1 again, the description of the configuration of the folding generation prediction system is continued. The prediction unit 4 outputs the prediction result obtained by the adaptive completion model 312 to the result display unit 5 . The result display unit 5 displays the prediction result to the operator in an easy-to-understand manner. By referring to the prediction result displayed on the result display unit 5 , the operator can perform an intervention operation for suppressing the occurrence of folds on the rolling stand to be predicted.

2. The processing flow of folding to generate predictions

FIG. 5 is a flowchart showing in detail the processing flow of the folding occurrence prediction system according to the present embodiment. In FIG. 5 , among the two flowcharts arranged on the left and right, the flowchart on the left shows the processing flow in the construction phase of the adaptive model, and the flowchart on the right shows the processing flow in the prediction phase.

First, the flow of processing in the construction phase of the adaptive model will be described according to the flowchart on the left. In the building phase of the adaptive model, step 101 is first performed. In step 101, the first data and the second data for constructing the adaptive model are collected and saved.

Here, a method for collecting the second data will be described in detail. For example, when collecting data for predicting the occurrence of folds in the sixth stand F6 of the finishing mill, data in the vicinity of the trailing end is collected for each of stands F1, F2, and F3 on the upstream side of stand F6. The data items of the acquired data are roll gap, roll shift amount, rolling load, thickness gauge thickness, and the like. Among them, the data of WS, DS, and the center part are collected about the roll gap, and the data of WS and DS are collected about the rolling load.

Collecting data near the trailing end means that enough time can be obtained before the prediction processing is performed and the operator is notified. For example, from the trailing end of the rolling material to the leading end for 30 seconds, and from here, it is collected upstream for 10 seconds. The data. In other words, among the 40 seconds of data near the end, the first 10 seconds of data are collected. In this case, an adaptive model was constructed using the 10 seconds of data collected. The time of 30 seconds is an example of a time approximately equal to the sum of the time for the prediction processing and notification to the operator, and the operator's preparation to avoid folding. By securing such a time, before the rolling material of the prediction target reaches the rolling stand of the prediction target, the presence or absence of the occurrence of folds in the rolling stand of the prediction target and all or a part of the occurrence location are predicted, Notifying the operator can prompt the operator to prepare to avoid folding.

In addition, according to the machine learning method and statistical method described later, better accuracy may be obtained without using all items of the second data collected or data collected at all times. Therefore, the folding occurrence prediction system of the present embodiment is configured to be able to appropriately select and use the required data from the data collected as the second data.

In step 102, it is determined whether the first predetermined number or more of data are stored or the second predetermined number or more of data are added. Here, the first predetermined number is an absolute number of data sufficient to be applied to machine learning and statistical methods. In the case of machine learning, it varies depending on the method, but generally requires 3,000 to 10,000 or more pieces of data. The second predetermined number is necessary for the determination for updating the adaptive model based on the newly added data because the data increases as the rolling progresses. It can be selected arbitrarily, but if the number is set to be small, it will be updated frequently, but the computational load will also increase. If the setting is large, the update frequency will be reduced, but there is a possibility that the new situation of rolling cannot be followed.

In step 101 , the collection and storage of the first data and the second data for constructing the adaptive model are continued until the conditions of step 102 are satisfied. Then, if the conditions of step 102 are satisfied, the flow proceeds to step 103 . Step 103 is the main process for building an adaptive model. Adaptive models implemented by methods such as NN, SOM, and ACC update their internal state through input data, enabling more accurate predictions.

In step 104, the adaptive model that has been constructed is saved as an adaptive completed model. If there is an existing adaptive completion model constructed using old data, the existing adaptive completion model is updated with the adaptive model constructed this time, and the updated adaptive completion model is saved.

Next, the processing flow of the prediction stage will be described according to the flowchart on the right. In the prediction stage, step 201 is first performed. In step 201, the rack number (k) of the rack to be folded to be predicted is set. Since folds are easily generated in the stands at the rear stage of the finishing mill, the stand numbers (k) may only be set to 4, 5, 6, and 7. In step 202, the rack number (k) is updated and reset one by one at each execution.

In step 203 , data for prediction is collected using the same collection method as the second data for adaptive model construction described in step 101 . The preceding rolling stand for which the prediction data is acquired needs to be determined in consideration of the time required for the prediction process, the time required to notify the operator via the display on the display device, and the time required for the operator to prepare for the avoidance of the folding operation. The following table shows the correspondence relationship between the rolling stand to be predicted (referred to as a prediction target stand in the table) and the previous rolling stand for which prediction data was collected (referred to as a predicted data acquisition stand in the table). one example.

Table 1

Prediction Object Rack Predictive Data Acquisition Rack F7 F1, F2, F3 or F1, F2 F6 F1, F2, F3 or F1, F2 F5 F1, F2 F4 F1, F2

In step 204, the adaptive completion model saved in step 104 is read, and prediction data is input to the adaptive completion model. For each prediction target rack, the adaptive completion model outputs the prediction result of the presence or absence of folding, and the prediction result of the generation location when folding occurs.

In step 205, it is checked whether or not prediction has been performed on all the prediction target racks. The processing from step 202 to step 204 is repeated until the prediction in all the prediction target racks is completed. Then, after the predictions in all the prediction target racks are completed, the process proceeds to step 206 to notify the operator of the prediction results. At this time, the operators in the operating room are often watching the actual rolling state on the opposite side of the glass window, or watching one of the television monitors installed in the operating room. Therefore, it is desired that the notification of the prediction result is also displayed in these two places where it is easy to observe. Here, the operator is notified of the prediction result after all the target racks are predicted, but the operator may be notified each time a fold is predicted to occur. That is, when the prediction is started from the m-th rack, if it is predicted that a fold occurs in the m-th rack, the situation may be notified immediately, and then the m+1-th rack may be predicted.

3. Adaptive model for predictions produced by folding

3-1.NN

In the simplest configuration, NN is a three-layer structure of input layer, intermediate layer, and output layer, and intermediate layers can also be added. In the case where the intermediate layer is composed of multiple layers, deep learning can also be performed. Each layer is composed of one or more neurons, and the neurons of each layer are combined with lines with weights. The output state of a neuron changes according to the level of the input value. When there is a teacher to learn, a so-called back-propagation method is generally used in which the output of the NN is compared with the teacher's signal, and the weight of the bonding line is updated in the reverse direction.

3-2.SOM

SOM does not require teacher data and only uses normal data, that is, data collected when no folds are generated. If the division of SOM is defined as 5×5, 10×10, 25×25, etc., each division becomes a neuron. Prepare a plane of the number of variables used in each neuron. FIG. 6 is a diagram illustrating a configuration example of an SOM and an outline of a prediction method using the configuration example of the SOM. In the configuration example of the SOM shown in FIG. 6 , the SOM is composed of 10×10 neurons.

A plane of the number of variables is prepared in each neuron. Here, the number of data items such as roll gap and rolling load included in the second data is assumed to be 20. A plane with 20 variables is prepared for each neuron. Specifically, a plane having an axis of the value of the variable and an axis of time is prepared. Each of the 20 variables has data corresponding to 10 seconds in the vicinity of the trailing end of the rolled material, as described in step 101 of the processing flow shown in FIG. 5 .

In the construction phase of the adaptive model, that is, the learning phase, a curve for 10 seconds on the initial plane in each neuron is randomly assigned. Of course, there are restrictions such as the non-repetition of each neuron. In addition, normal data of 10 seconds and 20 variables are stored as one set. The data is taken out in groups, and it is determined which neuron curve is closest to the entire data of the group. Then, let it belong to the neuron determined to be the closest. The same processing is performed for all groups of normal data, and finally the curve that becomes the center of gravity is determined by each variable in each neuron. Thus, the construction of the adaptive model is completed, that is, the learning is completed.

In the prediction phase, that is, the normal abnormality determination phase, each variable value of the prediction data, which is the determination target data, is compared with the barycentric value of each variable value in 100 neurons. Then, to which neuron the determination target data as a whole is close to is calculated, and the neuron determined to be the closest is selected. Next, calculate the distance between the curve of the center of gravity of each variable in the selected neuron and the value of each variable in the judgment object data, if the distance is a variable that greatly deviates from other distances (in the above example, the variable in 1), the data containing the variable is considered abnormal. That is, it is determined that folding has occurred.

The following table shows a verification example of the learning effect when an adaptive model is constructed using normal data by SOM, and abnormal data is detected using the adaptive completion model. In this verification example, among all 7650 data of the same steel type, 136 abnormal data with folded tails were included, but these abnormal data were detected 100%. By using this adaptive completion model, the presence or absence of the occurrence of folds can be predicted with an accuracy close to 100%.

Table 2

3-3.ACC

ACC basically applies the method of well-known control charts. The upper management limit (Upper Control Limit: hereinafter, referred to as UCL) and the lower management limit (lower control limit: hereinafter, referred to as LCL) are fixed in the management map, but ACC changes these according to the transition of data. Assuming that there is a certain time-series data and the period represents a transition of 10 seconds, for example, the aiming is shifted every 0.1 second, and 1 second is regarded as one section from here, and the UCL in each section is determined based on the standard deviation of each section. and LCL. At this time, if there is a degree of distortion in the data, UCL and LCL correction based on this is also performed. The method of this correction is described in "Betul Kan, and Berna Yazici" The Individuals Control Chart in Case of Non-Normality", Journal of Modern Applied Statistical Methods, Vol. 5, Issues 2, Article 28, Digital Commons@WayneState (2005)" middle.

Hereinafter, with reference to FIGS. 7 to 9 , the outline of the operation of the ACC will be described in more detail, taking the detection of tail folding as an example. At the stage of constructing the adaptive model, the normal data for 10 seconds near the tail end of the 20 selected variables is used as data for constructing the adaptive model in the same manner as the SOM. The 10 seconds in the vicinity of the tail end means 10 seconds in the vicinity of the front end from here, in addition to the 30 seconds in the vicinity of the rear end. In other words, it means the first 10 seconds of the 40 seconds near the end.

Data is collected as in the table shown in FIG. 7 . The column of the table is the volume number and the row data number. The number of roots of the volume is P roots. In the data of one volume, J is the first number of the 10-second data, and J+I is the last number of the 10-second data. Thus, 10 seconds of data includes one piece of data. Assuming that data is recorded every 0.1 second, in this case, I=10÷0.1=100 pieces.

In the table, a plurality of boxes are drawn in the row of the volume number p, and each box represents a window for calculating the standard deviation. The window for calculating the standard deviation was set to a 1-second interval, and the standard deviation of the data was calculated while shifting the window every 0.1 second. This calculation is performed for all 20 variables of the P normal volumes. Therefore, the standard deviation of each time period such as 0-1 second and 0.1-1.1 second can be P, and the distribution of the standard deviation can be formed. For example, the standard deviation in time period 1 is obtained as S[1, 1], S[2, 1], ..., S[P, 1]. By calculating the standard deviation σ[i] of the standard deviation S[1∼P,i] in the same time period i, the UCL can be taken to, for example, 3 times or 4 times the + side of σ[i], and the LCL can be Take, for example, 3 times or 4 times the - side of σ[i]. Thereby, UCL and LCL as shown in FIG. 8 can be obtained. However, FIG. 8 shows UCL and LCL for one variable, and UCL and LCL are calculated for each variable. In ACC, such determination of UCL and LCL corresponds to the construction of an adaptive model.

As shown in FIG. 9 , in the prediction stage, the data to be determined is compared with the UCL and LCL determined in the construction stage of the adaptive model. Then, using the sum of the number of data exceeding UCL or LCL and the distance from UCL or LCL as an evaluation value, the distance of the data to be judged from UCL and LCL is evaluated, and whether it is normal or abnormal is judged.

In NN, the abnormal data generated by folding and the normal data generated by no folding are input to construct a NN-based model. On the other hand, in SOM and ACC, only normal data without fold generation is input to construct a model, and when it deviates greatly from the model, it is determined that fold generation occurs. In general, the amount of rolled material that produces folds is much less than the amount of rolled material that does not produce folds. In addition, in the operation of the equipment, the normal operation time is also much longer than the abnormal state time. Therefore, in general, there is far less data representing abnormality than normality. In such a situation, the SOM and the ACC are often used to construct a model using the data of the normal state, and a structure that determines the abnormal state other than the normal state is more advantageous.

3-4. Characteristics of each adaptive model

The adaptive model is defined so that when data is input, the degree of relationship between internal components changes and the output changes. Examples of this are adaptive models constructed by various methods of NN, SOM, and ACC. In more detail, the suitable model suitable for the prediction produced by folding includes the suitable model defined by the following definition A, definition B, and definition C.

(A) When data is input, the degree of relationship between the internal components changes and the output changes, and after learning from a plurality of sets of inputs and outputs, the evaluation object inputs and outputs evaluation values

(B) When data is input, the degree of relationship between the internal components changes and the output changes, and after learning from a collection of multiple inputs, the evaluation object inputs and outputs the evaluation value

(C) When inputting data, the relationship between the internal components changes and the output changes, and after a statistical index is determined by a plurality of input sets, the input set is evaluated and the evaluation value is output

NN is an example of an adaptive model defined by Definition A. SOM is an example of an adaptive model defined by Definition B. ACC is an example of an adaptive model defined by definition C. In the table below, for each of NN, SOM, and ACC, the relationship to the definition of the adaptive model is shown.

table 3

4. Variation of the processing flow for prediction by folding

FIG. 10 is a flowchart showing a modification of the processing flow of the folding occurrence prediction system according to the present embodiment. The processing flow shown in FIG. 5 is different from the processing flow shown in FIG. 10 in the processing flow of the construction phase of the adaptive model. Specifically, the processing of step 104 in the processing flow shown in FIG. 5 is replaced by the processing of step 104 a in the processing flow shown in FIG. 10 . Although the adaptive model is updated in the adaptive model construction stage, step 104a is a step for confirming whether the prediction accuracy based on the updated result can be ensured. If the accuracy of the newly constructed adaptive model is sufficient, the adaptive completion model is updated and saved, otherwise it is not saved.

The method used to validate the constructed adaptive model is slightly different among the three methods above. In NN, data for learning is separated from data for validation. For example, assuming that there are 10,000 pieces of data, 500 or 1,000 pieces of data are randomly selected as data for verification, and the rest are data for learning. The verification data is input into the NN-based model that has learned the learning data, and it is verified whether or not the occurrence of folds is predicted with high accuracy. In SOM or ACC, since the model is built with normal data without folds, only abnormal data is used for validation. Input abnormal data into the model constructed by SOM or ACC, and if it is determined that there is folding, it is determined to be a high-precision model.

<Second Embodiment>

Next, the folding occurrence prediction system according to the second embodiment of the present invention will be described. In the present embodiment, after predicting the occurrence of folding, not only the prediction result is notified to the operator, but it is also used for the control of the entrance-side guide. The fold occurrence prediction system of the present embodiment includes a control device that operates the entrance-side guide plate of the rolling stand to be predicted when the prediction device predicts that folds will occur in the rolling stand to be predicted. The computer that functions as the control device may be a different computer from the computer that functions as the prediction device. In addition, a single computer may be made to function as a prediction device and also function as a control device by software.

FIG. 11 is a diagram showing an example of control of the entrance-side guides 31 and 32 using the folding prediction result. Here, an example of control will be described in the case where the rear end fold occurs in the k-th rack and the position is predicted to be DS. In the state shown in the upper stage of FIG. 11, the rolling material 100 is rolled by the k-1th stand. Snaking has not yet occurred. Then, it is assumed that the rolling material 100 zigzags after the time has elapsed and the state shown in the lower stage is reached, and the trailing end is separated from the k-1-th stand. An example of the predicted occurrence of end folds is when the rolled material 100 is in such a state. At this time, the control device (not shown) performs an operation to open the entry-side guide plate 31 of the DS in accordance with the passage of the trailing end in order to prevent the trailing end from hitting the entrance-side guide plate and breaking. In addition, when it cannot be determined which side of the WS and DS is folded, a control device (not shown) operates to open the inlet side guides 31 and 32 on both sides of the WS and DS in cooperation with the passage of the trailing end.

In the example shown in FIG. 11 , it is assumed that the tail end fold is generated, but the same control can be performed even when the front end fold is generated. That is, when it is predicted that the front end is folded, the control device (not shown) performs the operation of opening the entrance side guide in accordance with the passage of the front end. The side guide on the entrance side to be opened is the side guide on the side where the fold is generated. However, when it is not possible to determine whether the fold occurs at WS or DS, the operation to open the side guides on both sides of the WS and DS is performed.

By performing the above-described control on the entry-side guide plate of the rolling stand to be predicted, the operator can be assisted and stable work can be realized.

<Other Embodiments>

In the above-mentioned embodiment, the description was made with the object of the finishing rolling mill, but the present invention can also be applied to the rough rolling mill. In a rough rolling mill, reverse rolling can be achieved by repeating forward and reverse rolling a plurality of times. When the present invention is applied to a rough rolling mill, the rolling pass refers to each rolling, and the preceding rolling pass in the rolling sequence refers to the rolling performed before the last time.

In addition, in the above-mentioned embodiment, although the tail end folding has been described as a center, the present invention can be applied not only to the tail end folding, but also to the front end folding.

In addition, NN, SOM, and ACC have been described as examples of the adaptive model building method, but the adaptive model building method applicable to the present invention is not limited to these. For example, Random Forest (RF) using the idea of a search tree, Extra Trees, xgboost, etc., which are developed forms of RF, can also be applied.

Description of reference numerals

1: Adaptive model construction data collection and storage department

2: Adaptive Model Building Section

2a, 302: Adaptive model

3: Data collection section for forecasting

4: Forecasting Department

4a, 312: Adaptive Completion Model

5: Result display section

10: Prediction device

31, 32: Inlet side guide plate

100: Rolled material

Claims (10)

1. A fold occurrence prediction system for predicting the occurrence of a fold in hot rolling in which a plate-like metal material is heated to a high temperature and is rolled in a plurality of rolling passes, the fold being a phenomenon that the rolled material meanders or is bent in the width direction and occurs at the leading end or the trailing end of the rolled material,

the system is provided with one or more computers,

the one or more computers perform the following:

processing for collecting and storing data for adaptive model construction used in construction of an adaptive model for predicting the generation of a fold;

a process of constructing the adaptive model using the adaptive model construction data;

storing the self-adaptive model which is constructed, namely the self-adaptive completion model;

a process of collecting data for prediction used in prediction of generation of a fold; and

a process of predicting the generation of a fold by inputting the data for prediction to the adaptive completion model,

in the process of collecting and storing the data for adaptive model construction, a plurality of sets of first data and second data are collected as the data for adaptive model construction, the first data indicating the presence or absence of occurrence of a fold and a place of occurrence in a target pass to be predicted for the occurrence of a fold, the second data including information on a preceding pass when a rolled material associated with the first data is rolled in the preceding pass, the preceding pass being earlier in rolling order than the target pass, and an attribute related to the rolled material,

in the process of collecting the data for prediction, data including information on a preceding rolling pass in which a rolling material to be predicted is rolled in a preceding rolling pass preceding the rolling pass to be predicted and an attribute related to the rolling material are collected as the data for prediction,

in the process of predicting the occurrence of the fold, the presence or absence of the occurrence of the fold and all or a part of the occurrence location in the target rolling pass are predicted before the rolled material to be predicted reaches the target rolling pass.

2. The fold creation prediction system of claim 1,

the system is provided with a display device which is provided with a display,

the one or more computers are capable of,

processing for displaying a result of prediction of the generation of the fold on the display device is executed.

3. The fold creation prediction system of claim 1,

the one or more computers are capable of,

in a case where it is predicted that the fold is generated in the target rolling pass, a process of operating the entrance-side guide plate of the target rolling pass is performed.

4. The fold generation prediction system of claim 3,

the one or more computers are capable of,

in the process of operating the entrance-side guide, it is determined which end of the leading end and the trailing end of the rolled material to be predicted is folded, and the entrance-side guide is opened in accordance with the passage of the end on which the folding is generated.

5. The fold creation prediction system of claim 3,

the one or more computers are capable of,

in the process of operating the entrance-side guide, it is determined on which of the operating chamber side and the motor side of the subject rolling pass the fold is generated, and the entrance-side guide on which the fold is generated is opened.

6. The fold generation prediction system of claim 3,

the one or more computers are capable of,

in the process of operating the entrance-side guide, when it is not possible to determine which of the operating room side and the motor side of the target rolling pass is folded, the entrance-side guide on both the operating room side and the motor side is opened.

7. The fold creation prediction system of any of claims 1 to 6,

the one or more computers are capable of,

in the process of constructing the adaptive model, the adaptive model is constructed by machine learning or a statistical method that is included in the category of artificial intelligence, and the adaptive model is updated every time a certain number of data for constructing the adaptive model is newly obtained.

8. The fold creation prediction system of any of claims 1 to 6,

the one or more computers are capable of,

in the processing of collecting and storing the data for adaptive model construction, the presence or absence of occurrence of a fold and a location of occurrence in the target pass are determined by analyzing the image data of the rolled material that has passed through the target pass.

9. The fold creation prediction system of any of claims 1 to 6,

the one or more computers are capable of,

in the processing of collecting and storing the data for adaptive model construction, the presence or absence of the occurrence of the fold and the location of occurrence in the target rolling pass are determined based on the load applied to the entrance-side guide plate of the target rolling pass.

10. The fold creation prediction system of any of claims 1 to 6,

the one or more computers are capable of,

in the process of collecting and storing the data for adaptive model construction, the presence or absence of occurrence and the occurrence location of a fold in the target rolling pass, which are input by an operator via a human-machine interface HMI, are received.

Images